AI frontline guide: on December 18, NVIDIA Corp (NVDA.US) GTC China 2019 officially kicked off on the bank of Jinji Lake in Suzhou. NVIDIA Corp founder Huang Renxun wore that familiar black leather jacket to report on the achievements made by the company this year.

A year ago, the day before the opening of GTC China 2018 at the same venue, NVIDIA Corp's share price halved and its market value lost more than $20 billion. In the following year, NVIDIA Corp changed his former high-profile "arrogant" attitude, gathered his edge, and developed in a low profile, so that he did not hear NVIDIA Corp's "heroic words" for almost more than a year.

In the face of such a huge change in NVIDIA Corp's attitude, people can't help but ask: what happened to NVIDIA Corp this year?

Blockbuster technology releases, reviews new products and brings "little surprises"

"The more you buy,the more money you save!"

Yes, when you see this sentence, you should know: Huang Renxun is back, familiar leather jacket, familiar "plastic Chinese".

Today, Lao Huang has brought several major releases and new products in four areas:

Graphic technology

In terms of graphics technology, RTX has supported more games to optimize the light and shadow effects, including Microsoft Corp's "my World". In addition, NVIDIA Corp has reached a cooperation with Tencent START to provide players with cloud game services, even if the terminal performance is insufficient, it will not damage the game experience.

As the trump card of NVIDIA Corp's graphics technology, RTX has been adopted by the top rendering platform, and RTX Studio contains more than 40 products for content workers. In addition, Huang Renxun released OMNIVERSE, a large 3D graphics design platform for the construction industry, on the spot. It takes only 40 hours to render a scene that used to take 485 hours to complete.

HPC

Huang Renxun said that accelerated computing requires a complete stack tool, and excellent chips are just the starting point. Each application area requires a special mathematical kernel, algorithm library, and application software optimization. NVIDIA Corp CUDA platform has rich libraries, tools and applications. Last year alone, NVIDIA Corp released more than 500 SDK and libraries, both new and updated. Based on the continuous improvement of the software stack, the performance of users' existing GPU has also been improved, the deep learning training has been increased by 4 times in 3 years, and the deep learning reasoning has been improved by 2 times in 1 year.

Huang Renxun highlighted that Nvidia has added two new mainstream applications to CUDA: 5G vRAN and genome processing. And released a genome analysis toolkit Parabricks. This is a CUDA-accelerated genome processing toolkit that can be used to detect mutations and produce results consistent with recent industry standard GATK practices.

Note: the genome processing process is called GATk in the industry standard. This process can draw short strands of DNA according to references to reconstruct the genome and identify mutations, so it is a computing-intensive task.

The main features of PARABRICKS:

It can achieve 30,50x acceleration.

It can provide DeepVariant tools and use deep learning technology to detect gene variation. Based on the above advantages, some gene technology companies have adopted PARABRICKS, and with the help of several GUP servers, the genome can be processed according to the rate of data generated by its sequencer.

AI

At today's press conference, NVIDIA Corp launched TensorRT 7, the latest version of TensorRT. TensorRT 5, released last year, has been criticized by developers because it only supports CNN. In contrast, the latest version of TensorRT 7 can support various types of RNN, transformer, CNN, can integrate horizontal and vertical operations, design a large number of RNN configurations for developers to automatically generate code, point-by-point fusion of LSTM units, and even across multiple time steps for fusion, and as far as possible automatic low-precision processing. In addition, TensorRT 7 can support more than 1000 different computational transformations and optimizations. Huang Renxun said, for example, that with the support of TensorRT 7, it only takes 0.3 seconds to compile the session model on GPU.

The newly released robot platform chip Orin can handle higher precision sensor and sensing data, and has stronger security and prevention capabilities.

Orin chip

The chip consists of 17 billion transistors and integrates a new generation of NVIDIA GPU architecture and Arm Hercules CPU core as well as a new deep learning and computer vision accelerator. It can run 200 trillion calculations per second, almost seven times the performance of the previous generation of NVIDIA Xavier system-on-chip chips.

Orin can handle a large number of applications and deep neural networks running simultaneously in self-driving cars and robots, and meets system safety standards such as ISO 26262 ASIL-D.

As a software-defined platform, DRIVE AGX Orin can provide a compatible architecture platform for the development of fully self-driving vehicles from L2 to L5, helping OEM to develop large and complex software product lines. Because both Orin and Xavier can be programmed through open CUDA, TensorRT API and various libraries, developers can use cross-generational products after an one-time investment.

Robot

In the field of intelligent robots, NVIDIA Corp released ISAAC SDK, which provides a simulation and training platform, including a variety of pre-training models and a variety of reference application examples.

Left: simulated scene; right: robot is applied to real scene after learning

Through simulation, developers can test the robot in extreme situations (that is, difficult or abnormal situations) to further strengthen its training. These results are input into the training pipeline, so that the neural network can improve the accuracy based on real and simulated data.

Isaac SDK includes Isaac Robotics Engine (providing application framework), Isaac GEM (pre-built deep neural network models, algorithms, libraries, drivers, and API), reference applications for indoor logistics, and the first version of Isaac Sim (which provides navigation capabilities).

Use Isaac Sim to train robots and deploy the generated software to real robots running in the real world. This is expected to greatly speed up the development of the robot, so as to achieve comprehensive data training.

In the end, the release of an one more thing caught people's eyes:

The robot, called Leonardo, is the best embodiment of Isaac Sim's training results and can be grasped by simulated learning objects, Huang Renxun said. This is the embryonic form of the robot of the future.

Having seen the technical accumulation of NVIDIA Corp this year, I would like to talk about NVIDIA Corp's various experiences over the past year.

The demand for GPU slowed, the stock price plummeted three times, and NVIDIA Corp fell into a trough.

Seeing Boss Huang, who is still excited on stage today, it is hard to imagine that this is a person who has experienced three troughs in a year.

A year ago, the day before GTC China 2018, NVIDIA Corp's share price plummeted about 18% to $164.43 per share, and at one point in the following days it fell to $145.00. Although Huang Renxun at the press conference at that time tried his best to stir up the atmosphere at the scene, the act of canceling the interview after the meeting could still feel a trace of helplessness.

At that time, Huang Renxun said: the fall in share prices is directly related to the retreat of the cryptocurrency boom. In 2017-2018, the cryptocurrency went into a frenzy and the demand for computing power soared, causing the price of NVIDIA Corp GPU to rise, which deterred some buyers. As the boom recedes and the price of game cards falls back to the normal level, NVIDIA Corp naturally hopes that buyers who are blocked by the price will come back again to promote sales growth.

However, everything is not as smooth as expected. For a long time, NVIDIA Corp chip inventory is in a state of backlog, NVIDIA Corp CFO Colette Kress said in an interview: to completely solve the problem of excess GPU, it may take another 1 or 2 quarters.

Out of cautious consideration, on January 28, NVIDIA Corp issued a performance warning and lowered his revenue forecast for the fourth quarter by 500 million US dollars, directly causing NVIDIA Corp's share price to plunge more than 17% after the opening of the stock market, and finally closed at US $138.01, down 13.82%. The next day it closed at $131.6.

Some analysts believe that the slowdown in market demand for chips is still a major problem troubling NVIDIA Corp's development. Perhaps it was with this in mind that NVIDIA Corp bought Mellanox, an Israeli chip design company, for a record $6.9 billion in March, the largest acquisition in Yingweida history. Industry insiders believe that NVIDIA Corp's move is to get rid of the dependence on cryptocurrency and the game market, so as to concentrate on the business center on chip technology.

The money has been spent, but we still need to wait for the result. In May this year, NVIDIA Corp released his first-quarter results. According to the financial report, NVIDIA Corp's revenue in the first quarter of 2019 reached $2.22 billion, but it was down more than 30% from $3.007 billion in the same period last year. Net profit is even more worrying: net profit of $394 million is down by 68% compared with $1.244 billion in the same period last year.

One wave after another, before NVIDIA Corp had time to worry about the revenue of the next quarter, on June 18, the US government put five Chinese enterprises and institutions mainly developing chips and supercomputers on the "blacklist." this was followed by a collective collapse in the share prices of American chip companies, of which NVIDIA Corp closed at $151.76.

In August, NVIDIA Corp's second-quarter results showed that NVIDIA Corp's revenue in the quarter was $2.579 billion, down 30% from a year earlier and up 16% from $2.22 billion in the previous quarter. Net profit was $552 million, down 50% from $1.101 billion in the same period last year.

After two consecutive quarters of falling profits and three serious share price declines, NVIDIA Corp had an ugly transcript in the first half of 2019.

Low-key release of new products turn around and embrace overkill, can you survive?

After several setbacks, NVIDIA Corp became convergent and low-key, but also turned the bow to the chip, super-calculation-based business development.

Mellanox, which NVIDIA Corp acquired in March, is known for its InfiniBand interconnection technology, which, in conjunction with fast Ethernet products, can quickly transfer data through the system. In short, many small computers can be connected into mainframes. Huang Renxun has said: the acquisition of Mellanox aims to solve the problem of connecting a large number of fast computing nodes through an intelligent network structure to form the overall architecture of a huge data center-scale computing engine.

With Mellanox, NVIDIA Corp can provide customers with higher performance, higher utilization and lower operating costs. It is not difficult to see that when the development of the chip business is sluggish, NVIDIA Corp aims to overcalculate.

In June, NVIDIA Corp announced that he was working with Arm to develop a supercomputer, which confirmed this speculation. Ian Barker, vice president of NVIDIA Corp's accelerated computing department, said: Arm technology can provide an open architecture for supercomputing.

Also in June, NVIDIA Corp launched the 22nd supercomputer in the world: the DGX SuperPOD. According to NVIDIA Corp's official introduction, it has built-in 96 NVIDIA DGX-2H supercomputers and Mellanox interconnection technology in just three weeks, has 9.4 petaflops processing capacity, and can train a large number of deep neural networks needed for safe self-driving vehicles.

In November, at the SC19 supercomputing conference, NVIDIA Corp officially announced that Arm processors can use GPU acceleration technologies from NVIDIA Corp, such as CUDA. In addition, a new GPU product was also unveiled in a low profile-Tesla V100s, but according to media reports participating in the event, Huang Renxun did not mention this product in his speech that day.

With low-key new products and the determination to embrace overcalculation, NVIDIA Corp is looking for vitality in a desperate situation. On November 15, NVIDIA Corp released his third-quarter results, which showed that NVIDIA Corp's third-quarter revenue was $3.014 billion, down 5% from $3.181 billion in the same period last year, and its net profit was $899 million, down 27% from $1.23 billion in the same period last year.

Although the figures are still low compared with the same period last year, it is clear that the gap is narrowing-NVIDIA Corp is trying to recover.

Conclusion

Huang Renxun, wearing a black leather jacket, got rid of his fatigue and was energetic when he stood in the middle of the stage. What happened to NVIDIA Corp in 2019 will eventually pass with the passage of time. Although many problems, such as the chip business, remain to be solved, everything is taking a turn for the better, and perhaps soon, the "nuclear bomb plant" will burst into new power.